Recent predictions have been made of metallurgical properties of low-activation ferritic/martensitic steels alloys at the high irradiation levels (displacements per atom or dpa) needed for a fusion power plant as based on measurements at low irradiation levels where more data is available. These have been published for the yield stress and for the Charpy ductile to brittle transition temperature shift. The neural network model predictions use training data up to a certain dpa level to predict metallurgical properties above this level. This “extrapolation” mode of neural networks is explored in some detail. Our studies revealed an increasing accuracy of predictions as the test dpa level is increased for both yield stress and Charpy shift predictions. This result suggests that a model exists for these metallurgical properties as a function of dpa level which becomes more accurate as the available irradiation range in the training data is increased. The explanation suggested is that the metallurgical annealing which occurs as the irradiation level is increased simplifies the microstructure and makes prediction more reliable.